Automatic Speech Recognition•dzmsjs•6deditedExperience

A Brief Discussion on Methods for Reducing the Error Rate of Qwen ASR Models

There are currently many large models for automatic speech recognition on the market, and major companies have released their own models…… My experience: Qwen is the best to use. Overall, its error rate can be reduced to the lowest level, beating a host of other models. But how can you achieve a lower error rate than Qwen? My solution: multi-model fusion. In other words, when Qwen makes a mistake, other models have a chance of getting it right. Everyone's errors are all different; if they all get it wrong, then there's nothing we can do— they'll all be wrong. Based on this test result, fusing multiple models can ultimately compress the error rate to a minimum of 2.1776%, which is the bottleneck in this industry. An error rate below 4% is already extremely excellent. I'll share my specific code later. ---- [Oracle / Fusion Space] At least one model correct: 28004/28610 (97.88%) Majority of models correct: 26639/28610 (93.11%) All models incorrect: 606 All models made the same error: 393 Oracle CER lower bound: 2.1776%